<p>In recent years, virtual geographic environments have played a crucial role in enhancing public risk perception through disaster simulations. Previous studies, such as Zhu et al.¹, have proposed knowledge-driven visualization frameworks, offering valuable insights into public risk perception. Building on this foundation, this paper further focuses on the automated construction of knowledge organization and cross-platform adaptive presentation to better meet the public’s needs for understanding and participation in settings without professional guidance. This framework begins with a thorough analysis of the public’s specific needs for disaster visualization, using large language model (LLM) to extract triples and construct a detailed knowledge graph containing disaster geographic information. Under this knowledge framework, we realized precise modeling of virtual scenes and context-adaptive representation based on the theory of virtual geographic environments (VGEs). Then, we designed a cross-platform data organization and dynamic scheduling algorithm to enable content presentation across diverse devices. Finally, we conducted three groups of experiments using typical disaster cases, with user cognition assessed via eye-tracking. The experiment results indicate our method supports adaptive, smooth visualization on multiple platforms effectively. Compared to traditional approaches, it significantly improves debris flow disaster information dissemination efficiency by integrating scene context, user interest, demand response and spatial intelligence, offering advantages in standardized modeling, personalization and adaptive optimization, thereby enhancing public debris flow disaster information perception.</p>